Medication adherence is an important issue, as nonadherence all too frequently results in significant negative outcomes of interest to patients (e.g., hospitalization, seizures, loss of function and independence). Medication nonadherence is also largely preventable, and partly within the patient’s control. Unfortunately, studying nonadherence and its impact on outcomes can be quite difficult: not only do most medical studies require strict adherence, but studies that intentionally deprive patients of their treatments may be unethical and, besides, would not necessarily reflect the nonadherence patterns of real individuals. Fortunately, electronic health record (EHR) data, stored to keep track of an individual patient’s medical records, as a byproduct offers a way to study typical patients and the impact of their nonadherence on outcomes of interest to patients. EHR-based studies can even achieve a very large scale, up to millions of patients. However, adherence information in EHR data is largely stored in free text fields (i.e., in English as opposed to a structured database). This requires the development of a natural language processing (NLP) system to extract and represent the adherence-related information. This project will develop such an NLP tool based on state-of-the-art machine learning methods.
This adherence NLP tool will enable future large-scale studies on the relationship between nonadherence and patient-centered outcomes. The tool will not be limited to binary studies of adherence (e.g. does take medications versus does not), but will capture granular details (e.g., frequency, certainty, reasons for nonadherence) that will enable large “big data” studies to leverage their strength in scale to determine how precise nonadherence patterns impact outcomes.
The fundamental philosophy guiding this study is that in order to build a successful tool for extracting fine-grained adherence information from EHR notes, patients must be involved in the research process to craft a model of adherence. Simply put, in order to know what patterns of nonadherence to look for, one must engage with patients who understand the difficulties of adherence to treatment plans. We will convene an advisory group of patients, clinicians, and stakeholders with a focus on two chronic conditions known for nonadherence impacting patient-centered outcomes: diabetes and depression. These patient stakeholders will identify the most salient aspects of nonadherence, including those they typically communicate with their provider and those aspects they typically withhold. With an adherence model created with the help of patient stakeholders, the NLP tool created in this project will enable valuable research studies in this understudied area that nonetheless has enormous effects on the quality of patients’ lives.